Artificial Intelligence Software for Detecting Paroxysmal Atrial Fibrillation from Sinus Rhythm Monitor ECG: Development and Clinical Trial

Introduction: Detecting paroxysmal atrial fibrillation (pAF) from sinus rhythm could enable earlier intervention and stroke prevention. We developed a deep-learning Holter electrocardiograph (ECG) algorithm and prospectively evaluated its patient-level performance against 7-day AF outcomes. Methods: We curated 20,000 30-s sinus rhythm blocks (125 Hz) from Holter ECG data of patients with and without pAF, trained convolutional models with tenfold cross-validation, and assessed a separate validation set (n = 54; 27 pAF, 27 controls) to select the operating threshold. A multicenter prospective study then evaluated the algorithm using ten consecutive 30-s sinus rhythm blocks per patient with a 4/10 positive rule; patients with pAF underwent concurrent 7-day patch monitoring to anchor outcomes. Results: Cross-validation during development yielded mean sensitivity 84.2% and specificity 66.2%; the best tuned model achieved 84.9% sensitivity and 69.9% specificity on the separate set. In the clinical trial, among 24 patients with AF documented within 7 days and 20 controls, the device showed sensitivity 91.7% (95% confidence interval (CI) 73.0–99.0) and specificity 65.0% (40.8–84.6). No device-related adverse events occurred. Conclusion: An artificial intelligence (AI) analyzing short sinus rhythm Holter segments can identify patients who develop pAF within 7 days, supporting use as a triage tool for intensified rhythm monitoring. Trial Registration: UMIN-CTR UMIN000047182.